Published Wednesday, August 05, 2026 at 02:07 PM PT

Burbank · Wednesday, August 5, 2026 · 2:07 PM · 92°F, 43% humidity, wind 1 mph WSW (gusts 2), 29.35 inHg, UV 0, PM2.5 14

You think physics is about discovering nature’s laws. It isn’t. Physics is about replacing nature with a model you can simulate, interrogate, and optimize—and then discovering that simulating something you already “understand” is orders of magnitude more expensive than just observing it directly. This is the central, uncomfortable paradox of applied physics, and it’s woven through the entire history of the discipline.

We got here through a specific path. In the 1700s, French chemists and American gentlemen scientists spent decades arguing about electricity. Charles François du Fay found that gold leaf repels glass rubbed with silk but attracts amber rubbed with wool, and concluded that electricity came from two different fluids—vitreous and resinous—that could neutralize each other. What du Fay had actually discovered was polarity: the existence of two distinct charge states that behave as opposites. His two-fluid model wasn’t wrong, exactly; it was a description of an observable phenomenon, regularized enough to predict behavior in simple geometries. You could hold a rubbed rod near a piece of gold leaf, observe whether it was attracted or repelled, and map the pattern. But the two-fluid model gave you no insight into what was really happening, and it bought you no speed when you tried to predict behavior in complex situations. It was an organizing principle, not an explanation.

Ten years later, Benjamin Franklin showed up, dismissed the “two fluids” model, and proposed instead that electricity was a single fluid in either surplus (positive) or deficit (negative). He didn’t know which one was the surplus; he just guessed. It was a guess, but it was a productive guess, because it reduced apparent complexity to underlying simplicity. Two models → one model with two states. Suddenly you could write down a single principle: charge flows from regions of surplus to regions of deficit, and that flow was electricity. It wasn’t more accurate than du Fay’s model, not really. You could predict many of the same phenomena with both. But the single-fluid model was more economical. It suggested that underneath the apparent duality was a single underlying thing, and that suggestion proved generative. Every successful theory in physics works this way: it takes phenomena and asks, “What if all of this is actually just one thing, slightly different in configuration?”

By 1881, Hermann von Helmholtz had taken this further: both charges, he argued, were divisible into “elementary parts, each of which behaves like atoms of electricity.” George Johnstone Stoney estimated their value using Faraday’s laws. Stoney had a problem to solve: if electricity flowed, and flow meant something was moving, what was moving? What was the smallest unit? He could measure the total charge passed in electrolysis (Faraday had given him that). He could count the atoms dissolved (chemistry could do that). Divide the total charge by the number of atoms and you get the charge per atom. That charge, divided by some assumed number of charge carriers per atom, gave him a number. In 1894, Stoney coined the term “electron” for this “most remarkable fundamental unit of electricity.”

The pattern here is consistent and revealing: physicists spend enormous effort reducing apparent complexity to underlying simplicity. Two fluids → one fluid with two states → one fundamental particle. Each step feels like progress. Each step feels like understanding. And it is understanding—but understanding the what, not the how much or the where or the when.

Consider what Stoney’s electron allowed: it meant you could write down equations for how charge moves through matter. It meant you could predict whether a given configuration would conduct or insulate. It meant you could calculate the force between two charged particles. But understanding that electricity comes from elementary particles tells you nothing about how to route current through a wire in microseconds without inducing crosstalk from adjacent traces. It tells you nothing about predicting how a lightning bolt will branch as it descends through air. It tells you nothing about whether a given semiconductor layout will have acceptable electromagnetic interference. The fundamental unit of electricity is known to better than one part in a billion. The behavior of electrical systems in complex geometries remains hard. Knowledge of the microscopic and prediction of the macroscopic are not the same thing.

Which brings us to a man shooting golf balls into a hole while a computer suffers.

Stuff Made Here (the YouTube channel) has been documenting a project in which they built a golf-ball-putting robot. The physics involved is textbook high-school mechanics: a rigid body (the club) collides with another rigid body (the ball), transferring momentum under known laws (Newton’s), and the ball then travels through space subject to gravity and air resistance, bouncing and rolling until it (hopefully) lands in a hole. Every single law involved has been known for centuries. The mathematics is not exotic. A motivated high schooler could write the equations on a whiteboard.

But here’s where it gets interesting: simulating where the ball goes for a specific hit is fast. You give the computer a club angle, a swing speed, a club-ball contact duration, and the coefficient of friction between club and ball, and the ball’s moment of inertia. The computer solves the collision equations, computes the initial velocity, the spin rate, and the launch angle. Then it propagates the ball forward through space, accounting for gravity (downward acceleration of 32 feet per second squared), air resistance (which depends on velocity and the ball’s cross-sectional area), and the Magnus effect (the ball curves because it’s spinning). The ball bounces off the ground, losing energy to friction and deformation. It rolls, slowing due to rolling resistance. Eventually it stops. All of this can be computed in milliseconds. You can simulate a thousand shots per second on a modern computer.

But determining which angle and speed will make it go in—that is a different problem entirely. You cannot solve that in the few milliseconds available while the club is swinging. You cannot run the inverse calculation in real time. The ball’s final resting place depends on angle, speed, and spin rate in a complex, nonlinear way. Small changes in any of those parameters produce small changes in where the ball ends up—except when they don’t. A change of half a degree in launch angle might mean the ball misses the hole by an inch, or by a foot, depending on where the hole is and what the terrain looks like. There is no simple closed-form solution. You cannot write down an equation that says “if you swing at this angle with this speed, the ball will go in.” You have to simulate.

So when the ball is placed on the course, the computer performs what Stuff Made Here describes as “an enormous amount of pre-computation.” It simulates all the possible hits the club could make at all speeds and angles. Angles from, say, -5 degrees to +5 degrees of the measured aim point. Speeds from, say, 80% to 120% of nominal. That’s 1000 angles × 100 speed variations = 100,000 simulations. Each simulation takes milliseconds, but a hundred thousand of them takes seconds. So the robot simulates not just one course configuration but multiple trajectories through it, capturing how a ball curves, where it bounces, where it ends up. It records which ones score. “It’s kind of like when Doctor Strange sits down to consider all the potential futures to find the one where the Avengers succeed,” the creator says. Except this took 15 to 20 hours initially. Optimized, it still takes about an hour. For each ball placement. Every time the ball moves.

Let me be direct about what this means: you have completely understood the physics of a golf ball. You have written down the exact equations. You have measured every parameter—friction coefficients, air density, spin rates, club dynamics, bounce properties. You have validated your model against reality by placing the ball on the course, swinging the robot, and comparing where the ball actually landed to where the model predicted. You have understanding that would have seemed like sorcery to Newton. And yet, predicting behavior in a new configuration requires you to run millions of simulations—burning CPU cycles like gasoline—just to find the answer you already “know” exists. The laws of physics do not make prediction cheap. They make it possible, which is not the same thing.

This is not accidental. This is the design of physics itself.

Look at the chronological record of atmospheric science. In 1751, Charles Le Roy recognized dew point—the temperature at which air becomes saturated and water begins to condense. In 1782, Jacques Charles flew a hydrogen balloon to measure temperature and pressure at altitude, demonstrating that temperature and pressure both decrease with height. In 1805, Pierre Simon Laplace developed his law of pressure variation with height, showing that the pressure at any altitude follows an exponential decay. In 1933, Tor Bergeron published his foundational paper describing how precipitation forms from supercooled water droplets, explaining the mechanism of cloud microphysics. In 1986, K. Emanuel successfully modeled tropical cyclones as Carnot heat engines, showing that a hurricane’s intensity is fundamentally limited by thermodynamic principles derived from how much heat it can extract from warm ocean water and dump into the upper atmosphere. Over two and a half centuries, physicists incrementally constructed a model of the atmosphere—layer by layer, mechanism by mechanism, principle by principle.

And yet: weather prediction beyond two weeks remains nearly impossible. Not because we don’t understand the physics. We do, reasonably well. The problem is that the atmosphere is chaotic. Edward Lorenz discovered this in the 1960s when he tried to run a weather simulation twice with nearly identical initial conditions. The second run differed from the first not by a small amount, but drastically—complete divergence after a few simulated days. He had changed the initial condition by 0.000127 instead of 0.000128—one part in ten thousand. That tiny change cascaded into large changes in the predicted outcome. This is sensitive dependence on initial conditions, the signature of chaos. Predicting whether it will rain on August 15th in Los Angeles is not a matter of understanding the governing equations—it is a matter of measuring infinitely precise initial conditions, which you don’t have, and then simulating a coupled system of billions of variables for 360 hours of simulated time on a computer powerful enough to handle it. The computational requirement is ferocious: meteorological centers around the world run supercomputers continuously, each solving billions of equations per forecast, and still the forecast is only reliable five to seven days out.

So physicists did what physicists always do: they built a different model. Instead of trying to predict where a specific air parcel will be, they ask what is the probability distribution of where it might be. This is the random walk. This is diffusion. When Brownian motion was first observed—a pollen grain jiggling randomly in water—physicists could write down the individual equations of motion for the pollen grain being struck by individual water molecules. But there are 10^23 water molecules in a glass of water. You cannot simulate all of them. And you don’t need to. The random collisions average out to a pattern: the distance a diffusing particle moves is proportional to the square root of time: x_rms = b√N, where N is the number of steps and b is the step length. This is not a simplification that emerges naturally from first principles by setting up the collision equations and solving them. It is a model—a replacement for first-principles calculation—because first-principles calculation is too expensive.

George Johnstone Stoney, who had estimated the electron by dividing Faraday’s laws by atomic counts, spent no time calculating the trajectory of a single electron in a conductor. He knew it couldn’t be done. Instead, when physicists try to predict the motion of ions in an electrolyte, or the path of a particle in a gas, they use the diffusion equation. It gives you a bulk description—a density function that evolves over time—rather than individual particle trajectories. It works. It produces predictions that match reality. But it is a surrender. It says: I will give up trying to track individuals and instead tell you about the population. I will give up real-time prediction and instead give you a probability distribution.

This is the core insight: physics is not about discovering nature’s laws. It is about building simulacra—models whose behavior closely enough mimics nature to allow prediction. And the deeper you understand the underlying laws, the more apparent it becomes that prediction is expensive. Understanding the rules of chess does not let you play blindfolded in real time. A computer can calculate all possible moves from a given position (there are roughly 30 legal moves on average), then calculate all possible replies to each of those, and so on, to a depth of 20 or 30 moves ahead. But calculating 30^20 possibilities is not feasible. So chess engines instead use heuristics: they evaluate a position using pattern recognition and material counting, not by brute-force lookahead. They sacrifice accuracy for speed. Understanding electricity does not let you route signals through arbitrary geometries in nanoseconds without crosstalk and signal degradation. Understanding the rules of fluid flow does not let you predict turbulence in your shower with real-time accuracy.

The history of physics is therefore not the history of discovering truth. It is the history of building progressively more detailed simulacra, each one more expensive to run than the last. We went from “electricity comes from two fluids” to “electricity comes from the motion of electrons” to “electrons move as described by quantum mechanics” to “quantum mechanics in a conductor under a temperature gradient with defects at the grain boundaries looks like this”—and each step gave us deeper understanding but also more expensive simulation. We went from “things fall at a constant acceleration” to “things fall under gravity, but air resistance matters” to “air resistance is turbulent and depends on Reynolds number” to “turbulent flows exhibit fractal structure at multiple scales and their energy dissipates according to the Kolmogorov cascade”—and again, deeper understanding, more expensive prediction.

The creator of the golf-putting robot knows this intimately. They built a computer that understands physics better than most humans ever will. They measured every parameter with precision instruments. They wrote the equations from Newton’s laws and collision mechanics. They validated against reality by testing the robot against the course. And when they place a ball down, they must choose: either spend an hour computing where to aim the club, or don’t know where to aim the club. There is no third option. There is no way to skip the simulation. There is no clever insight that lets you look at a golf course and immediately know how to hit a putt without computation. The laws of physics are deterministic, but deterministic does not mean fast. Deterministic means predictable; it does not mean cheaply predictable.

This reveals something uncomfortable about how we think about physics. We imagine that understanding a system means you can predict it quickly. We imagine that reducing a phenomenon to its fundamental components gives you leverage to compute the result efficiently. Both are sometimes true. But just as often, they’re not. The fundamental particles of electricity are known to better than one part in a billion. The geometry of the golf course is precisely scanned with millimeter accuracy. The equations of motion are textbook, validated in countless experiments. And yet: wait an hour, or don’t play the shot.

Physics doesn’t describe nature in the sense of encoding its behavior in language that you can execute quickly. It replaces nature with a formal system—a set of rules and parameters—that we can interrogate computationally. That system can be right (that is, it can accurately predict real outcomes with high fidelity) without being cheap to query. A map that perfectly captures terrain is useless if you cannot afford to look at it fast enough to navigate. A model that perfectly captures physics is useless if you cannot afford to run it in real time. The map and the physics are both tools for prediction, and both hit the same wall: the question you can afford to ask is limited by the time you have to compute the answer.

The practical path forward—the one that people like the creator of this golf robot have taken—is to abandon real-time prediction altogether. Instead, you precompute. You simulate all possible futures (or at least, a large representative sample of them). You store the results. When you need to act, you look it up. Doctor Strange had to evaluate 14 million possible futures to find the one where the Avengers win. The golf robot had to evaluate millions of possible shots to find the ones where the ball goes in. In both cases, the only way to get a good answer in time was to pay the computational price upfront, in parallel, on a machine powerful enough that the calculation finishes within acceptable bounds. A chess engine precomputes endgame tables—all possible positions with a small number of pieces remaining—so that when it reaches the endgame, it can look up the winning move instantly instead of calculating it. A structural engineer runs finite-element simulations of thousands of possible bridge designs before deciding which one to build, because running those simulations is cheaper than building bridges and watching them fail. The precomputation model is ubiquitous in engineering because it is the only model that works at the scale we operate at.

This is not a temporary limitation that will be solved by faster computers. Doubling computing speed just means you can simulate twice as much—which lets you ask twice as hard a problem, which makes the simulation twice as expensive again. It’s an arms race between predictive ambition and computational budget. The only way to break the cycle is to change the question: not “where will the ball go?” but “where might the ball go?” The random walk replaces deterministic prediction with probabilistic description. Diffusion replaces particle tracking with a density function. Weather forecasting replaces exact prediction with ensemble forecasts—a range of possible outcomes weighted by probability. All of these are surrenders—acknowledgments that real-time prediction is not available, so we’ll give you what we can afford to give you instead. And all of them have proven remarkably useful, not because they are as accurate as exact prediction (they’re not), but because they are accurate enough and cheap enough.

The history of physics is therefore the history of increasingly sophisticated surrender. We surrender the hope that we can predict individual electrons, so we describe their statistical behavior and the collective phenomena that emerge. We surrender the hope that we can predict the exact location of a raindrop, so we describe probability distributions and ensemble averages. We surrender the hope that we can aim a golf club in real time by thinking, so we precompute millions of shots and pick the best one. We surrender the hope that we can calculate stress distribution in a complex aerospace composite by hand, so we discretize the problem, run it on a computer, and get an answer good enough that the plane doesn’t fall out of the sky. At each stage, we call it progress. At each stage, it is progress. But progress in physics is not progress toward omniscience. It is progress toward having better models—models that are more accurate but also more expensive, more intricate, more demanding of computational resources.

The uncomfortable truth is that understanding and predicting are not the same. You can understand something completely and still be unable to predict it quickly. The golf-putting robot understands the physics of a golf ball at a level that would have seemed like wizardry two centuries ago. It measures things that Newton couldn’t measure. It calculates things that would have taken a mathematician weeks by hand. And yet it must wait an hour. That hour is not a bug in the model. It is the price of the model’s accuracy. It is the tax you pay for understanding something well enough to simulate it at all.

This is where physics ends and engineering begins. Physics gives you the laws. It gives you a way to replace nature with a formal system you can interrogate. Engineering finds ways to live within computational budgets. Engineering builds precomputation strategies, trades off accuracy against speed, finds clever ways to avoid simulation altogether when possible (measure the real thing instead), and when simulation is unavoidable, figures out how to make it fast enough. And somewhere between the two, in that hour-long pause before the robot swings the club, all of applied science lives.